How to Use Seedance 2.5 (2026): Prompts, Timestamps, Cost

Seedance 2.5 is ByteDance’s 30-second video model, and it changes enough of the prompting contract that a Seedance 2.0 prompt is no longer the right prompt. Timestamps work now. The reference budget went from 15 assets to 50. The task type is inferred from words in your prompt, and getting those words wrong returns the wrong kind of video rather than an error.

Model ID:        bytedance/seedance-2.5 on a gateway, dreamina-seedance-2-5-260628 on ByteDance's API
Clip length:     4 to 30 seconds, or -1 to let the model choose
Resolution:      480p and 720p today; vendor says 1080p starts 2026-08-17; 4K stays on 2.0
Audio:           synchronized, on by default, suppressible per element
References:      50 assets max, as 30 images + 10 videos + 10 audio clips
New in 2.5:      integer-second timestamps, audio-only references, mov output, 11 languages
Cost:            $0.11/s at 480p, $0.24/s at 720p
Measured:        6s clip billed $0.66 in 103s; 30s clip billed $3.30 in 200s
Snapshot:        2026-08-14

Everything below is checked against the official Seedance 2.5 tutorial and prompt guide published by ByteDance, plus three real generations we ran through bytedance/seedance-2.5 on 2026-08-13. Where our runs disagree with the documentation, we say so.

What Changed From Seedance 2.0?

Four things, and the vendor lists them itself. The 2.5 prompt guide carries a short section titled “Differences from Seedance 2.0” that is worth more than any of the launch coverage:

To that, the capabilities table in the tutorial adds the numbers people quote: 30 seconds instead of 15, 50 reference assets instead of 15, and audio-only references, which the 2.0 series refuses without an accompanying image or video.

One inconsistency to know about before you check our work: that same capabilities table lists an identical set of output aspect ratios for both models, 21:9 through 9:16 plus adaptive. The flexible-ratio claim in the row above is the prompt guide’s, and it is about what the input assets can make the output become, not about a longer dropdown.

What nobody in the launch coverage leads with is the direction the resolution went. Seedance 2.5 tops out at 720p today; the tutorial page, last updated 2026-08-13, still says 1080p and 4K are unsupported. Part of that gap now has a date on it. The ModelArk pricing page carries a notice that Seedance 2.5 “will officially support 1080p output starting from August 17, 2026, Beijing time”, with its own token rate and a discount attached (both below). 4K has no such notice, so Seedance 2.0 stays the only model in the family that reaches it. Until the 17th, anyone delivering a master file is on 2.0 regardless. We put the whole trade in Seedance 2.5 vs 2.0, including what each one costs for the same clip.

Where Can You Run Seedance 2.5?

Two routes that bill you honestly, and a consumer app that does not publish its rates.

The resource-pack gate is the part that surprises people. ByteDance’s own documentation is explicit that Seedance models activate only while you hold a matching prepaid pack, so the “just add a card and call the API” mental model from other vendors does not apply. Packs expire and are not refundable.

Through a gateway the model is one string. On ofox it is bytedance/seedance-2.5 on the same /v1/videos endpoint as the 2.0 tiers, so switching between them is a one-word diff. The request and polling shape has not changed since we documented it in the Seedance 2.0 API access guide, and the free and access question has the same answer it did for 2.0: there is no free API tier on any route.

How Do You Write a Seedance 2.5 Prompt?

Three blocks, in this order: a one-sentence summary, a beat-by-beat plot, then the details that must hold across the whole clip. That is the structure ByteDance’s guide asks for, and it maps onto how the model reads your text far better than the paragraph of adjectives most video prompts turn into.

The published shape:

One-sentence summary:  subject + location + event + genre/style + camera movement
Detailed plot:         timestamps or Shot N, each with visuals, camera, action, dialogue, sound
Additional notes:      what stays constant throughout: angle, environment, atmosphere, audio

Two rules travel with it. Write positively, because negative constraints are only supported for subtitles and audio (“No subtitles”, “No BGM; generate only environmental sounds and action sounds”). And bind every reference asset explicitly in the text, numbered by upload order, rather than relying on labels drawn inside the image.

ByteDance also ships a prompt-engineering skill for the model, which is the single most useful thing in the guide that nobody has written about:

npx --yes skills@latest add \
  "https://arkdocs-en.tos-ap-southeast-1.volces.com/skills/" \
  --skill sd25-pe \
  --yes

We ran that command to check it resolves: it writes .agents/skills/sd25-pe/SKILL.md and registers the skill with whichever agent CLIs it finds, Claude Code and Codex among them. Then /sd25-pe plus your rough prompt rewrites it into the structure above.

Do Timestamps Actually Work Now?

Yes, and this is the rule that reversed. The Seedance 2.0 prompt guide tells you not to do this: “The model’s support for precise timing (such as 0–3 seconds) is unstable, and forcibly limiting duration may lead to abnormal generation results.” Our Seedance 2.0 guide passes that advice on, because it was correct. For 2.5 the same vendor writes the opposite: timestamps are supported at integer-second granularity, with three accepted styles.

Three constraints come with it. Keep the timeline continuous, because gaps like “0-3s… 5-6s…” are called out as a mistake. Budget roughly one beat every two to three seconds, since too little plot in a window makes the model improvise and too much makes it cut frantically or drop plot. And do not use timestamps for high-frequency actions; the guide’s own counter-example is “shake your head three times per second.”

We tested the claim rather than repeating it. One prompt, three beats, explicit second ranges:

0-2s: the paper fox sits still, then its ears twitch twice.
2-4s: the fox unfolds into a single flat sheet of orange paper.
4-6s: the flat sheet folds itself back into the fox, which hops once
      toward the camera and freezes.

At 480p, seed fixed, generate_audio: true. Sampling the returned clip at t=1s, t=3s and t=5s puts the fox intact, mid-unfold, and airborne mid-hop respectively, which is the script. The frame-by-frame comparison against Seedance 2.0 on the identical prompt is in the companion piece; the short version is that both models hit the beats, and 2.5 lands them closer to the second and actually performs the hop.

Stretching a script is where it gets interesting. We reran the same three-beat prompt at duration: 30 without adding plot. The result is coherent for the full 30 seconds, holds the subject, and never cuts, but it spends about ten seconds on a flat sheet of paper because there is nothing else in the script to do. The documentation predicts exactly this. A 30-second clip wants roughly ten beats, not three stretched.

How Many Reference Assets Should You Attach?

The ceiling is 50. The recommended working range is nowhere near it.

The 2.0 guidance to stay well under the asset limit has not been repealed, it has been rescaled. What genuinely changed is that audio can now stand alone as a reference, so voice or music can drive a generation with no image attached, and multi-view subject images are supported instead of discouraged. For more than five subjects the guide still asks for single-view images, split one per view rather than one image containing several views.

One practical note for anyone calling through a gateway rather than ByteDance directly: the reference ceiling you can actually use is whatever your gateway passes through, and those have historically lagged the model. Test with two reference images before you design a pipeline around thirty.

Which Task Type Are You Actually Running?

Seedance 2.5 sorts every request into locked and unlocked tasks, and locked tasks refuse parameters you would normally set. Seedance 2.0 makes no such distinction, which is why prompts ported straight across start throwing asynchronous errors.

Intent is read from your words. Editing needs one of edit video, add, insert, remove, delete, modify, replace, change to. Extension needs one of extend forward, extend backward, continue, continue from, extend the story. Miss them and your edit request is classified as a reference job, which returns a new video built loosely from your clip rather than your clip with one thing changed.

The escape hatch is omni_reference_task_type. Set it to edit or extend and the API validates the locked-task constraints at submit time, so you get a synchronous error instead of paying for a wrong-shaped result. Leave it at auto and validation happens after the task starts. The model still re-derives the task type from your prompt afterwards, so the parameter narrows the failure window rather than closing it.

For editing and extension, set output_format to mov. The guide recommends it for both, and mov is the format that carries the yuv444p chroma and PCM audio that make a re-edit look like the same footage.

What Does Seedance 2.5 Cost?

$0.11 per second of finished video at 480p, $0.24 at 720p, and video input costs more than text or images.

Real clips, real invoices, both from 2026-08-13:

Both land on exactly $0.11 per second, so the published rate is the rate, and a 30-second clip does not cost 30 seconds of premium.

ByteDance bills the same model differently: for 480p and 720p output, dreamina-seedance-2-5-260628 is $10.70 per million tokens with no video input and $6.40 with video, and tokens are estimated as (input video duration + output duration) × width × height × frame rate / 1024. Our 6-second clip came back 854×480 at 24 fps, which works out to about 57,600 tokens, or roughly $0.62. The per-second rate we actually paid is within about 7% of ByteDance’s own math for the identical clip, and it skips the prepaid pack. That is the honest way to compare a gateway to source: run the vendor’s formula on a clip you already generated.

The 1080p tier that arrives on 2026-08-17 is priced separately and higher: $11.70 per million tokens with no video input and $7.00 with video, listed at 72% of that through 2026-09-17 at 14:00 Beijing time, which ByteDance summarises as “starting at approximately USD 0.41 per second”. Run the token formula on both and that discounted 1080p second is about four times a 480p second at source, because token count scales with pixel area. It is a tier, not a rounding difference.

Two more billing details worth knowing. Failed generations are not charged, on either route, including failures from content moderation. And promotions still favour the older family for anything you can generate today: ByteDance is discounting the two cheap 2.0 tiers through 2026-09-07, taking Mini to 40% of list and Fast to 75% while the base 2.0 model stays at $7.00 per million tokens, and ofox shows 10% off the 2.0 rates against a $0.07 and $0.16 list. The one Seedance 2.5 promotion on the board applies to 1080p output only, and the pricing page says in the same sentence that 480p and 720p are not eligible. So every resolution 2.5 can currently produce is at full price.

How Do You Call Seedance 2.5 From Code?

Submit, poll, download. The endpoint is asynchronous, and a 30-second clip takes minutes rather than seconds, so build for polling or hand it a callback_url from the start.

import os, time, requests

BASE = "https://api.ofox.run/v1"
H = {"Authorization": f"Bearer {os.environ['OFOX_API_KEY']}"}

body = {
    "model": "bytedance/seedance-2.5",
    "prompt": (
        "3D animated style, bright clean studio lighting, a small orange origami "
        "paper fox on a white paper desk. "
        "0-2s: the paper fox sits still, then its ears twitch twice. "
        "2-4s: the fox unfolds into a single flat sheet of orange paper. "
        "4-6s: the flat sheet folds itself back into the fox, which hops once "
        "toward the camera and freezes. "
        "Fixed camera, medium close-up. No subtitles. "
        "Sound: soft paper rustling on each fold, one gentle chime on the freeze."
    ),
    "duration": 6,
    "resolution": "480p",
    "aspect_ratio": "16:9",
    "generate_audio": True,
}

job = requests.post(f"{BASE}/videos", json=body, headers=H).json()
while True:
    s = requests.get(f"{BASE}/videos/{job['id']}", headers=H).json()
    if s["status"] in ("completed", "failed"):
        break
    time.sleep(5)

print(s["status"], s.get("usage"))

That is the exact request that produced the frames above, minus the seed. usage comes back as {"video_seconds": 6, "video_cost": "0.6600000000"}, which is the number to log if you are tracking spend per clip rather than trusting an estimate.

Two failures are worth catching by hand, because both are cheap to hit and both return before generation starts:

{"error":{"code":"unsupported_parameter",
  "message":"resolution \"1080p\" not supported; allowed: [480p 720p]"}}

{"error":{"code":"unsupported_parameter",
  "message":"duration 30 out of range [4, 15]"}}

The first is Seedance 2.5 refusing 1080p. The second is Seedance 2.0 refusing a 30-second request, which is the error you will hit the first time you point old code at the new duration or new code at the old model.

What Should You Do First?

Generate one 4-second 480p clip, at $0.44, before you plan anything. It tells you whether your subject survives content moderation, whether audio comes back the way you want, and what your gateway actually passes through, and it costs less than a coffee.

Then scale in this order: get the beats right at 6 seconds with timestamps, add references one at a time rather than in a batch of thirty, and only move to 720p once the script is settled, because that is the point where the per-second rate more than doubles. If your delivery format needs 1080p before 2026-08-17, or 4K at all, stop here and read the comparison with Seedance 2.0 instead, because no amount of prompting gets today’s 2.5 past 720p.

References

• Dreamina Seedance 2.5 tutorial
• Dreamina Seedance 2.5 prompt guide
• Dreamina Seedance 2.0 series prompt guide
• ModelArk pricing
• ModelArk resource packs
• ofox model page: Seedance 2.5
• ofox model page: Seedance 2.0
• ofox video API reference